It seems the Manchester website is down. Link to the paper: https://github.com/jjfumero/jjfumero.github.io/blob/master/files/papers/2021/vee21-GPUMemoryJIT.pdf

It seems the Manchester website is down. Link to the paper: https://github.com/jjfumero/jjfumero.github.io/blob/master/files/papers/2021/vee21-GPUMemoryJIT.pdf https://github.com/jjfumero/jjfumero.github.io/blob/master/f... Just to complement the previous answer. It varies from kernel to kernel. Regarding reductions, the strategy implemented in the TornadoVM JIT compiler allows us to execute reductions offloaded from Java sequential code within 85% of the hand-written OpenCL Kernels: https://www.researchgate.net/publication/327871451_Using_Compiler_Snippets_to_Exploit_Parallelism_on_Heterogeneous_Hardware_A_Java_Reduction_Case_Study https://www.researchgate.net/publication/327871451_Using_Com...